Combine Latent Noise

Every unsampling workflow has one boring node that does the real math, and in logtd/ComfyUI-InversedNoise that node is Combine Latent Noise. It's not flashy - no sliders, no toggles - but it's the piece that keeps the inversion honest, which is why it's worth understanding even if you never wire it yourself.

What it does

CombineNoiseLatentNode takes three inputs - latent, noise, and sigmas - and returns one LATENT named noise. The math is a single line: it subtracts the clean latent (scaled by the first sigma) from the noisy latent. What's left over is the noise component that was added to the image in the first place.

That's the core trick of inversion: a noisy latent isn't a mystery blob, it's "image signal + noise scaled by the schedule." If you know the schedule's first sigma, you can solve for the noise exactly instead of guessing. This node does that solve, then hands you the recovered noise for re-injection later in the graph.

In practice the decomposition looks like:

• latent - the clean, VAE-encoded image
• noise - the same latent after noise was added
• sigmas - the schedule from a BasicScheduler, which tells it how much noise the top of the schedule implies
• output noise - the peeled-off noise term, ready to wire into your sampling step

It's bookkeeping, but it's the load-bearing kind: without it you'd be re-encoding and losing precision, or hand-waving the noise and watching structure drift.

When you'd use it

Honestly? Inside the pack's intended inversion graph, and not much outside it. If you're building a custom-sampling pipeline that recovers, edits, then re-injects noise, this is the node that separates "what the image is" from "what was added to it." If you're doing ordinary txt2img or img2img, you don't need it - and the pack's other nodes (the Inversed Euler Sampler for the recovery, Mix Noise with Latent for the creative slack) are the parts that actually change your output.

Installing it

Same as the rest of the pack - no dependencies beyond stock ComfyUI, no model files. Via ComfyUI Manager search "InversedNoise", or:

cd ComfyUI/custom_nodes
git clone https://github.com/logtd/ComfyUI-InversedNoise
# restart ComfyUI

Gotchas

The one thing that trips people up: it requires a real sigmas input, so it only works inside a custom-sampling graph - you can't just drop it between two latents and get a number. And since the whole pack is Experimental/WIP, treat the output as approximate. If the recovered noise looks noisy-but-wrong, the issue is usually the inversion quality upstream, not this node's arithmetic.